path: root/open_issues/dbus.mdwn

[[!meta copyright="Copyright © 2011, 2012, 2013, 2014 Free Software Foundation,
Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

[[!tag open_issue_hurd]]

The dbus problems are due to missing scm credentials [[sendmsg_scm_creds]] and socket credentials
[[pflocal_socket_credentials_for_local_sockets]]. There was also a problem with short timeout in 
[[select]], but that has been solved in Debian by setting a minimum timeout of 1ms.

[[!toc]]


# IRC, freenode, #hurd, 2011-11-26

    <antrik> BTW, how much effort is necessary to fix dbus?
    <pinotree> basically, have pflocal know who's the sender
      (pid/uid/gid/groups) in the socket send op


# IRC, freenode, #hurd, 2011-12-16

    <braunr> pinotree: what's the problem with dbus ?
    <pinotree> braunr: select() returning 0 changed fd's with very short (eg <
      1ms) timeouts when there are actually events;

[[select]].

    <pinotree> and missing socket credentials

[[sendmsg_scm_creds]].

    <braunr> oh
    <braunr> which socket creds interface ?
    <pinotree> bsd, i.e. with SCM_CREDENTIALS payload for cmsg on
      {recv,send}msg()
    <braunr> ok
    <braunr> SCM_RIGHTS too ?
    <braunr> the select issue seems weird though
    <pinotree> hm no, that's for passing fd's to other processes
    <braunr> is it specific to pflocal or does dbus use pfinet too ?
    <pinotree> iirc on very short timeouts the application has no time waiting
      for the reply of servers
    <braunr> i see
    <pinotree> braunr: http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=79358
    <braunr> thanks
    <pinotree> (the interesting messages are from #53 and on)
    <braunr> 2000 eh ... :)
    <braunr> hm i agree with neal, i don't understand why the timeout is given
      to the kernel as part of the mach_msg call


# IRC, freenode, #hurd, 2011-12-20

    <braunr> hm, i don't see any occurrence of SCM_CREDENTIALS in dbus
    <braunr> only SCM_RIGHTS
    <pinotree> braunr: yes, that one
    <braunr> oh
    <braunr> i thought you said the opposite last time
    <pinotree> dbus/dbus-sysdeps-unix.c, write_credentials_byte and
      _dbus_read_credentials_socket (with #define HAVE_CMSGCRED)
    <braunr> hm
    <braunr> which version ?
    <braunr> i don't see anything in 1.4.16
    <pinotree> 1.4.16
    <braunr> grmbl
    <braunr> ah, i see
    <braunr> SCM_CREDS
    <pinotree> if you want, i have a simplier .c source with it
    <braunr> no i'm just gathering info
    <pinotree> ok
    <braunr> what's the deal with SCM_CREDS and SCM_CREDENTIALS ?
    <braunr> bsd vs sysv ?
    <braunr> oh, http://lists.debian.org/debian-hurd/2002/03/msg00135.html
    <braunr> so we actually do want both SCM_CREDS and SCM_RIGHTS for debus
    <braunr> dbus
    <pinotree> SCM_RIGHTS is a different matter, it is for passing fd's
    <braunr> yes
    <braunr> but it's used by dbus
    <braunr> so if we can get it, it should help too
    <pinotree> there's a preliminary patch for it done by emilio time ago, and
      iirc it's applied to debian's glibc
    <braunr> ah, he changed the libc
    <braunr> right, that's the only sane way
    <pinotree> iirc roland didn't like one or more parts of it (but i could be
      wrong)
    <braunr> ok


# IRC, freenode, #hurd, 2013-07-17

    <teythoon> btw pinotree, what happened to your efforts to make dbus work?
    <pinotree> not much, my initial patch was just a crude hack, a better
      solution requires more thinkering and work 
    <teythoon> yes, ive seen that
    <teythoon> but that was only a tiny patch against the libc, surely there
      must be more to that?
    <pinotree> not really
    <teythoon> and the proper fix is to patch pflocal to query the auth server
      and add the credentials?
    <pinotree> possibly
    <teythoon> that doesn't sound to bad, did you give it a try?
    <pinotree> not really, got caught in other stuff


# IRC, freenode, #hurd, 2013-09-02

    <gnu_srs1> something is wrong with libc0.3 since the switch to 2.17. dbus
      does not run any longer when rebuilt
    <gnu_srs1> the latest build of dbus was with 2.13: libc0.3-dev: already
      installed (2.13-38)
    <pinotree> debug it
    <gnu_srs1> Yes, I will. Maybe somebody could rebuild it and verify my
      findings?
    <pinotree> gnu_srs1: your finding is "doesn't work", which is generic and
      does not help without investigation
    <gnu_srs1> just rebuild it and: e.g. ./build-debug/bus/dbus-daemon --system
      (--nofork)
    <pinotree> gnu_srs1: please, debug it
    <gnu_srs1> I have partially already. But maybe the problems only shows on
      my box. I'll rebuild on another box before continuing debugging.
    <pinotree> gnu_srs1: are you, by chance, running a libc or something else
      with your scm_creds work?
    <gnu_srs1> I did, but I've backed to 2.17-92 right now.
    <gnu_srs1> sane problem with dbus on another box, something's fishy:-(
    <gnu_srs1> braunr: any good way to find out if the dbus problems are with
      libpthread? Setting a breakpoint with libc0.3-dbg installed.
    <braunr> gnu_srs1: i don't know

See [[glibc]], *Missing interfaces, amongst many more*, *`SOCK_CLOEXEC`*.


# IRC, freenode, #hurd, 2013-09-04

    <gnu_srs> Hi, looks like dbus requires abstract socket namespace: #undef
      HAVE_ABSTRACT_SOCKETS What's missing?
    <pinotree> uh?
    <pinotree> abstract unix sockets are a Linux feature, and surely it is not
      mandatory for dbus
    <gnu_srs> Looks like dbus exits if they are not supported:
    <gnu_srs>  dbus_set_error (error, DBUS_ERROR_NOT_SUPPORTED,  "Operating
      system does not support abstract socket namespace\n");        _dbus_close
      (listen_fd, NULL);  1061       return -1;
    <pinotree> that is enclosed in a if (abstract)
    <pinotree> and that parameter is set to true in other places (eg
      dbus/dbus-server-unix.c) only when HAVE_ABSTRACT_SOCKETS is defined
    <pinotree> so no, abstract sockets are not mandatory
    <gnu_srs> Well this code is executed e.g. when running emacs remotely in
      X. Have to dig deeper then to see why.
    <pinotree> maybe it could have to do the fact that your dbus server is
      running in linux and runs by default using such sockets type
    <pinotree> but yes, you need to dig better
    <gnu_srs> pinotree: You are right. when running natively the problem is:
    <pinotree> *drums*
    <gnu_srs> Manually: Process /usr/lib/at-spi2-core/at-spi-bus-launcher
      exited with status 1
    <pinotree> eh?
    <gnu_srs> Error retrieving accessibility bus address:
      org.freedesktop.DBus.Error.Spawn.ChildExited: ^
    <pinotree> most probably that service does not start due to the lack of
      socket credentials which affects dbus
    <pinotree> uninstall or disable those additional services, they are not
      your problem
    <gnu_srs> credentials is enabled. which services to remove?
    <pinotree> dunno


# IRC, freenode, #hurd, 2013-09-11

    <gnu_srs> Hi, looks like frebsd had (2008) the same problem as hurd when
      sending credentials over PF_INET: 
    <gnu_srs>
      http://lists.freebsd.org/pipermail/freebsd-hackers/2008-May/024577.html
    <gnu_srs> Since the dbus code is about the same now (2013), maybe they
      added support?
    <gnu_srs> The next message in the thread confirms that the dbus code is
      invalid, does anybody have pointers?
    <pinotree> from what i've seen so far, socket credentials are done only for
      local sockets (ie PF_UNIX)
    <pinotree> i don't see how things like uid/gid/pid of the socket endpoint
      can have anything to do with AF_INET
    <pinotree> and socket credentials in dbus are used only in the [local]
      socket transport, so there's no issue


# IRC, freenode, #hurd, 2013-09-12

    <gnu_srs> pinotree: Yes, there is an issue with dbus and AF_INET, see
      test/corrupt.c: tests /corrupt/tcp and /corrupt/byte-order/tcp:-/
    <pinotree> gnu_srs: what's wrong with those? they are just testing the
      connection over a tcp socket
    <pinotree> as said above, socket credentials shouldn't be used in such
      cases
    <gnu_srs> They are, see also test/relay.c: /relay and /limit tests:-(
    <pinotree> how are they?
    <pinotree> please be more specifc...
    <gnu_srs> Just run the tests yourself with DBUS_VERBOSE=1
    <pinotree> you are claiming there is a problem, so please specify what is
      the actual issue
    <gnu_srs>  DBUS_VERBOSE=1 build-debug/test/test-relay 
    <pinotree> you are claiming there is a problem, so please specify what is
      the actual issue
    <gnu_srs> same with test-corrupt
    <gnu_srs> look at the verbose output:  Failed to write credentials: Failed
      to write credentials byte: Invalid argument
    <gnu_srs> coming from pfinet since PF_INET is used.
    <pinotree> check what it does on linux then
    <pinotree> put an abort() at the start of the read/write socket credential
      functions in dbus-sysdeps-unix.c and see whether it is triggered also on
      linux
    <gnu_srs> SO_PEERCRED is used for linux and LOCAL_CREDS is used for
      kfreebsd, so we are on our own here:-/
    <pinotree> and linux' SO_PEERCRED works also on AF_INET sockets? i'd doubt
      it
    <gnu_srs>
      http://stackoverflow.com/questions/10037086/so-peercred-vs-scm-credentials-why-there-are-both-of-them
    <pinotree> yes, i know the difference, but please read what i asked again
    <gnu_srs> I'll check to be sure...
    <braunr> gnu_srs: user credentials are not supposed to be passed through an
      AF_INET socket
    <braunr> how hard is that to understand ?
    <gnu_srs> OK, linux use send since CMSGCREDS is not defined to write
      credentials. Working on how they are received.
    <gnu_srs> braunr: I do understand, but the dbus code tries to do that for
      Hurd:-(
    <pinotree> then it should do that on linux too
    <pinotree> (since the local socket credentials code is isolated in own
      functions, and they are called only for the unix transport)
    <gnu_srs> Happiness:-D, almost all dbus tests pass!
    <gnu_srs> 17(17) dbus tests pass:)
    <braunr> gnu_srs: hopefully your patch does things right
    <gnu_srs> which patch
    <braunr> adding credentials through unix socket
    <braunr> isn't that what you're doing ?
    <gnu_srs> the mail to MLs is from the stock installed packages.
    <braunr> ?
    <gnu_srs> the test reports are with the SCM_CREDS patches, but I stumbled
      on the SCM_RIGHTS issues reported to MLs
    <gnu_srs> no patches applied, just test the attached file yourself.
    <braunr> so what's your work about ?
    <gnu_srs> I'm working on SCM_CREDS, yes, and created patches for dbus,
      glib2.0 and libc.
    <gnu_srs> the mail was about some bug in the call to io_restrict_auth in
      sendmsg.c: without any of my patches applied (another image) 
    <teythoon> gnu_srs: you have to give us more context, how are we supposed
      to know how to find this sendmsg.c file?
    <pinotree> (it's in glibc, but otherwise the remark is valid)
    <pinotree> s/otherwise/anyway/


# Emails

# IRC, freenode, #hurd, 2013-10-16

    <braunr> gnu_srs: how could you fail to understand credentials need to be
      checked ?
    <gnu_srs> braunr: If data is sent via sendmsg, no problem, right?
    <braunr> gnu_srs: that's irrelevant
    <gnu_srs> It's just to move the check to the receive side.
    <braunr> and that is the whole problem
    <braunr> it's not "just" doing it
    <braunr> first, do you know what the receive side is ?
    <braunr> do you know what it can be ?
    <braunr> do you know where the corresponding source code is to be found ?
    <gnu_srs> please, describe a scenario where receiving faulty ancillary data
      could be a problem instead
    <braunr> dbus
    <braunr> a user starting privileged stuff although he's not part of a
      privileged group of users for example
    <braunr> gnome, kde and others use dbus to pass user ids around
    <braunr> if you can't rely on these ids being correct, you can compromise
      the whole system
    <braunr> because dbus runs as root and can give root privileges
    <braunr> or maybe not root, i don't remember but a system user probably
    <pinotree> "messagebus"
    <gnu_srs> k!
    <braunr> see http://www.gnu.org/software/hurd/open_issues/dbus.html
    <braunr> IRC, freenode, #hurd, 2013-07-17
    <braunr> <teythoon> and the proper fix is to patch pflocal to query the
      auth server and add the credentials?
    <braunr> <pinotree> possibly
    <braunr> <teythoon> that doesn't sound to bad, did you give it a try?


# IRC, freenode, #hurd, 2013-10-22

    <gnu_srs> I think I have a solution on the receive side for SCM_CREDS :)

    <gnu_srs> A question related to SCM_CREDS: dbus use a zero data byte to get
      credentials sent.
    <gnu_srs> however, kfreebsd does not care which data (and credentials) is
      sent, they report the credentials anyway
    <gnu_srs> should the hurd implementation do the same as kfreebsd?
    <youpi> gnu_srs: I'm not sure to understand: what happens on linux then?
    <youpi> does it see zero data byte as being bogus, and refuse to send the
      creds?
    <gnu_srs> linux is also transparent, it sends the credentials independent
      of the data (but data has to be non-null)
    <youpi> ok
    <youpi> anyway, what the sending application writes does not matter indeed
    <youpi> so we can just ignore that
    <youpi> and have creds sent anyway
    <braunr> i think the interface normally requires at least a byte of data
      for ancilliary data
    <youpi> possibly, yes
    <braunr>        To pass file descriptors or credentials over a SOCK_STREAM,
      you need to  send  or
    <braunr>        receive  at  least  one  byte  of  non-ancillary  data  in
      the same sendmsg(2) or
    <braunr>        recvmsg(2) call.
    <braunr> but that may simply be linux specific
    <braunr> gnu_srs: how do you plan on implementing right checking ?
    <gnu_srs> Yes, data has to be sent, at least one byte, I was asking about
      e.g. sending an integer 
    <braunr> just send a zero
    <braunr> well
    <braunr> dbus already does that
    <braunr> just don't change anything
    <braunr> let applications pass the data they want
    <braunr> the socket interface already deals with port rights correctly
    <braunr> what you need to do is make sure the rights received match the
      credentials
    <gnu_srs> The question is to special case on a zero byte, and forbid
      anything else, or allow any data.
    <braunr> why would you forbid 
    <braunr> ?
    <gnu_srs> linux and kfreebsd does not special case on a received zero byte
    <braunr> same question, why would you want to do that ?
    <gnu_srs> linux sends credentials data even if no SCM_CREDENTIALS structure
      is created, kfreebsd don't
    <braunr> i doubt that
    <gnu_srs> To be specific:msgh.msg_control = NULL; msgh.msg_controllen = 0;
    <braunr> bbl
    <gnu_srs> see the test code:
      http://lists.debian.org/debian-hurd/2013/08/msg00091.html
    <braunr> back
    <braunr> why would the hurd include groups when sending a zero byte, but
      only uid when not ?
    <gnu_srs> ?
    <braunr> 1) Sent credentials are correct:
    <braunr> no flags: Hurd: OK, only sent ids
    <braunr> -z Hurd: OK, sent IDs + groups
    <braunr> and how can it send more than one uid and gid ?
    <braunr> "sent credentials are not honoured, received ones are created"
    <gnu_srs> Sorry, the implementation is changed by now. And I don't special
      case on a zero byte.
    <braunr> what does this mean ?
    <braunr> then why give me that link ?
    <gnu_srs> The code still applies for Linux and kFreeBSD. 
    <gnu_srs> It means that whatever you send, the kernel emits does not read
      that data: see
    <gnu_srs> socket.h: before  struct cmsgcred: the sender's structure is
      ignored ...
    <braunr> do you mean receiving on a socket can succeed with gaining
      credentials, although the sender sent wrong ones ?
    <gnu_srs> Looks like it. I don't have a kfreebsd image available right now.
    <gnu_srs> linux returns EPERM
    <braunr> anyway
    <braunr> how do you plan to implement credential checking ?
    <gnu_srs> I'll mail patches RSN


# IRC, freenode, #hurd, 2013-11-03

    <gnu_srs> Finally, SCM_CREDS (IDs) works:) I was on the right track all the
      time, it was just a small misunderstanding.
    <gnu_srs> remains to solve the PID check
    <youpi> gnu_srs: it should be a matter of adding
      proc_user/server_authenticate
    <gnu_srs> there are no proc_user/server_authenticate RPCs?
    <gnu_srs> do you mean adding them to process.defs (and implement them)?
    <youpi> gnu_srs: I mean that, yes


# IRC, freenode, #hurd, 2013-11-13

    <gnu_srs> BTW: I have to modify the SCM_RIGHTS patch to work together with
      SCM_CREDS, OK?
    <youpi> probably
    <youpi> depends on what you change of course


# IRC, freenode, #hurd, 2013-11-15

    <gnu_srs> Hi, any ideas where this originates, gdb? warning: Error setting
      exception port for process 9070: (ipc/send) invalid destination port
    <braunr> gnu_srs: what's process 9070 ?
    <gnu_srs> braunr: It's a test program for sending credentials over a
      socket. Have to create a reproducible case, it's intermittent.
    <gnu_srs> The error happens when running through gdb and the sending
      program is chrooted:
    <gnu_srs> -rwsr-sr-x 1 root root 21156 Nov 15 15:12
      scm_rights+creds_send.chroot


## IRC, freenode, #hurd, 2013-11-16

    <gnu_srs> Hi, I have a problem debugging a suid program, see
      http://paste.debian.net/66171/
    <gnu_srs> I think this reveals a gnumach/hurd bug, it makes things behave
      strangely for other programs.
    <gnu_srs> How to get further on with this?
    <gnu_srs> Or can't I debug a suid program as non-root?
    <pochu> gnu_srs: if gdb doesn't work for setuid programs on hurd, I suppose
      you could chmod -s the binary you're trying to debug, login as root and
      run it under gdb
    <gnu_srs> pochu: When logged in as root the program works, independent of
      the s flag setting.
    <pochu> right, probably the setuid has no effect in that case because your
      effective uid is already fine
    <pochu> so you don't hit the gdb bug in that case
    <pochu> (just guessing)
    <gnu_srs> It doesn't work in Linux either, so it might be futile.
    <gnu_srs> trying
    <pochu> hmm that may be the expected behaviour. after all, gdb needs to be
      priviledged to debug priviledged processes
    <gnu_srs> Problem is that it was just the suid properties I wanted to
      test:(
    <braunr> gnu_srs: imagine if you could just alter the code or data of any
      suid program just because you're debugging it


## IRC, freenode, #hurd, 2013-11-18

    <gnu_srs> Hi, is the code path different for a suid program compared to run
      as root?
    <gnu_srs> Combined with LD_PRELOAD?
    <teythoon> gnu_srs: afaik LD_PRELOAD is ignored by suid programs for
      obvious security reasons
    <gnu_srs> aha, thanks:-/
    <braunr> gnu_srs: what's your problem with suid ?
    <gnu_srs> I made changes to libc and tried them out with
      LD_PRELOAD=... test_progam. It worked as any user (including root),
    <gnu_srs> but not with suid settings. Justus explained why not.
    <braunr> well i did too
    <braunr> but is that all ?
    <braunr> i mean, why did you test with suid programs in the first place ?
    <gnu_srs> to get different euid and egid numbers

    <gnu_srs> hi, anybody seen this with eglibc-2.17-96: locale: relocation
      error: locale: symbol errno,
    <gnu_srs> version GLIBC_PRIVATE not defined in file libc.so.0.3 with link
      time reference
    <teythoon> yes, I have
    <teythoon> but afaics nothing did break, so I ignored it


## IRC, freenode, #hurd, 2013-11-23

    <gnu_srs> Finally 8-)
    <gnu_srs> Good news: soon both SCM_CREDS _and_ SCM_RIGHTS is supported
      jointly. RFCs will be sent soon.


## IRC, freenode, #hurd, 2013-12-05

    <gnu_srs> I have a problem with the SCM_CREDS patch and dbus. gamin and my
      test code runs fine.
    <gnu_srs> the problem with the dbus code is that it won't work well with 
    <gnu_srs> auth_user_authenticate in sendmsg and auth_server_authenticate in
      recvmsg.
    <gnu_srs> Should I try to modify the dbus code to make it work?
    <youpi> unless you manage to prove that dbus is not following the posix
      standard, there is no reason why you should have to modify dbus
    <gnu_srs> I think the implementation is correct,
    <gnu_srs> but auth_user_authenticate hangs sendmsg until
      auth_seerver_authenticate is executed in recvmsg.
    <gnu_srs> and dbus is not doing that, so it hangs in sendmsg writing a
      credentials byte.
    <gnu_srs> well the credentials byte is definitely non-posix.
    <gnu_srs> I found a bug related to the HURD_DPORT_USE macro too:-(
    <youpi> ah, yes, auth_user_authenticate might be synchronous indeed, let me
      think about it
    <gnu_srs> Nevertheless, I think it's time to publish the code so it can be
      commented on:-D
    <youpi> sure
    <youpi> publish early, publish often


# IRC, freenode, #hurd, 2014-01-17

    <gnu_srs> youpi: as a start all our requested dbus changes are now
      committed, and in Debian unstable
    <youpi> good :)


# IRC, freenode, #hurd, 2014-01-30

    <pochu> dbus has some known problems
    <pere> known fixes too?
    <pochu> http://www.gnu.org/software/hurd/open_issues/dbus.html
    <gnu_srs> pochu: Maybe that page should be updated:
      http://lists.nongnu.org/archive/html/bug-hurd/2013-12/msg00150.html
    <youpi> gnu_srs: well, maybe you can do it :
    <youpi> )